mm/hugetlb.c

   1 /*
   2  * Generic hugetlb support.
   3  * (C) William Irwin, April 2004
   4  */
   5 #include <linux/gfp.h>
   6 #include <linux/list.h>
   7 #include <linux/init.h>
   8 #include <linux/module.h>
   9 #include <linux/mm.h>
  10 #include <linux/sysctl.h>
  11 #include <linux/highmem.h>
  12 #include <linux/nodemask.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/mempolicy.h>
  15 #include <linux/cpuset.h>
  16 #include <linux/mutex.h>
  17
  18 #include <asm/page.h>
  19 #include <asm/pgtable.h>
  20
  21 #include <linux/hugetlb.h>
  22 #include "internal.h"
  23
  24 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
  25 static unsigned long nr_huge_pages, free_huge_pages, resv_huge_pages;
  26 unsigned long max_huge_pages;
  27 static struct list_head hugepage_freelists[MAX_NUMNODES];
  28 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
  29 static unsigned int free_huge_pages_node[MAX_NUMNODES];
  30 /*
  31  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
  32  */
  33 static DEFINE_SPINLOCK(hugetlb_lock);
  34
  35 static void clear_huge_page(struct page *page, unsigned long addr)
  36 {
  37         int i;
  38
  39         might_sleep();
  40         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
  41                 cond_resched();
  42                 clear_user_highpage(page + i, addr);
  43         }
  44 }
  45
  46 static void copy_huge_page(struct page *dst, struct page *src,
  47                            unsigned long addr)
  48 {
  49         int i;
  50
  51         might_sleep();
  52         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
  53                 cond_resched();
  54                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE);
  55         }
  56 }
  57
  58 static void enqueue_huge_page(struct page *page)
  59 {
  60         int nid = page_to_nid(page);
  61         list_add(&page->lru, &hugepage_freelists[nid]);
  62         free_huge_pages++;
  63         free_huge_pages_node[nid]++;
  64 }
  65
  66 static struct page *dequeue_huge_page(struct vm_area_struct *vma,
  67                                 unsigned long address)
  68 {
  69         int nid = numa_node_id();
  70         struct page *page = NULL;
  71         struct zonelist *zonelist = huge_zonelist(vma, address);
  72         struct zone **z;
  73
  74         for (z = zonelist->zones; *z; z++) {
  75                 nid = zone_to_nid(*z);
  76                 if (cpuset_zone_allowed(*z, GFP_HIGHUSER) &&
  77                     !list_empty(&hugepage_freelists[nid]))
  78                         break;
  79         }
  80
  81         if (*z) {
  82                 page = list_entry(hugepage_freelists[nid].next,
  83                                   struct page, lru);
  84                 list_del(&page->lru);
  85                 free_huge_pages--;
  86                 free_huge_pages_node[nid]--;
  87         }
  88         return page;
  89 }
  90
  91 static void free_huge_page(struct page *page)
  92 {
  93         BUG_ON(page_count(page));
  94
  95         INIT_LIST_HEAD(&page->lru);
  96
  97         spin_lock(&hugetlb_lock);
  98         enqueue_huge_page(page);
  99         spin_unlock(&hugetlb_lock);
 100 }
 101
 102 static int alloc_fresh_huge_page(void)
 103 {
 104         static int nid = 0;
 105         struct page *page;
 106         page = alloc_pages_node(nid, GFP_HIGHUSER|__GFP_COMP|__GFP_NOWARN,
 107                                         HUGETLB_PAGE_ORDER);
 108         nid = next_node(nid, node_online_map);
 109         if (nid == MAX_NUMNODES)
 110                 nid = first_node(node_online_map);
 111         if (page) {
 112                 page[1].lru.next = (void *)free_huge_page;      /* dtor */
 113                 spin_lock(&hugetlb_lock);
 114                 nr_huge_pages++;
 115                 nr_huge_pages_node[page_to_nid(page)]++;
 116                 spin_unlock(&hugetlb_lock);
 117                 put_page(page); /* free it into the hugepage allocator */
 118                 return 1;
 119         }
 120         return 0;
 121 }
 122
 123 static struct page *alloc_huge_page(struct vm_area_struct *vma,
 124                                     unsigned long addr)
 125 {
 126         struct page *page;
 127
 128         spin_lock(&hugetlb_lock);
 129         if (vma->vm_flags & VM_MAYSHARE)
 130                 resv_huge_pages--;
 131         else if (free_huge_pages <= resv_huge_pages)
 132                 goto fail;
 133
 134         page = dequeue_huge_page(vma, addr);
 135         if (!page)
 136                 goto fail;
 137
 138         spin_unlock(&hugetlb_lock);
 139         set_page_refcounted(page);
 140         return page;
 141
 142 fail:
 143         spin_unlock(&hugetlb_lock);
 144         return NULL;
 145 }
 146
 147 static int __init hugetlb_init(void)
 148 {
 149         unsigned long i;
 150
 151         if (HPAGE_SHIFT == 0)
 152                 return 0;
 153
 154         for (i = 0; i < MAX_NUMNODES; ++i)
 155                 INIT_LIST_HEAD(&hugepage_freelists[i]);
 156
 157         for (i = 0; i < max_huge_pages; ++i) {
 158                 if (!alloc_fresh_huge_page())
 159                         break;
 160         }
 161         max_huge_pages = free_huge_pages = nr_huge_pages = i;
 162         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
 163         return 0;
 164 }
 165 module_init(hugetlb_init);
 166
 167 static int __init hugetlb_setup(char *s)
 168 {
 169         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
 170                 max_huge_pages = 0;
 171         return 1;
 172 }
 173 __setup("hugepages=", hugetlb_setup);
 174
 175 #ifdef CONFIG_SYSCTL
 176 static void update_and_free_page(struct page *page)
 177 {
 178         int i;
 179         nr_huge_pages--;
 180         nr_huge_pages_node[page_to_nid(page)]--;
 181         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
 182                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
 183                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
 184                                 1 << PG_private | 1<< PG_writeback);
 185         }
 186         page[1].lru.next = NULL;
 187         set_page_refcounted(page);
 188         __free_pages(page, HUGETLB_PAGE_ORDER);
 189 }
 190
 191 #ifdef CONFIG_HIGHMEM
 192 static void try_to_free_low(unsigned long count)
 193 {
 194         int i;
 195
 196         for (i = 0; i < MAX_NUMNODES; ++i) {
 197                 struct page *page, *next;
 198                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
 199                         if (PageHighMem(page))
 200                                 continue;
 201                         list_del(&page->lru);
 202                         update_and_free_page(page);
 203                         free_huge_pages--;
 204                         free_huge_pages_node[page_to_nid(page)]--;
 205                         if (count >= nr_huge_pages)
 206                                 return;
 207                 }
 208         }
 209 }
 210 #else
 211 static inline void try_to_free_low(unsigned long count)
 212 {
 213 }
 214 #endif
 215
 216 static unsigned long set_max_huge_pages(unsigned long count)
 217 {
 218         while (count > nr_huge_pages) {
 219                 if (!alloc_fresh_huge_page())
 220                         return nr_huge_pages;
 221         }
 222         if (count >= nr_huge_pages)
 223                 return nr_huge_pages;
 224
 225         spin_lock(&hugetlb_lock);
 226         count = max(count, resv_huge_pages);
 227         try_to_free_low(count);
 228         while (count < nr_huge_pages) {
 229                 struct page *page = dequeue_huge_page(NULL, 0);
 230                 if (!page)
 231                         break;
 232                 update_and_free_page(page);
 233         }
 234         spin_unlock(&hugetlb_lock);
 235         return nr_huge_pages;
 236 }
 237
 238 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
 239                            struct file *file, void __user *buffer,
 240                            size_t *length, loff_t *ppos)
 241 {
 242         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
 243         max_huge_pages = set_max_huge_pages(max_huge_pages);
 244         return 0;
 245 }
 246 #endif /* CONFIG_SYSCTL */
 247
 248 int hugetlb_report_meminfo(char *buf)
 249 {
 250         return sprintf(buf,
 251                         "HugePages_Total: %5lu\n"
 252                         "HugePages_Free:  %5lu\n"
 253                         "HugePages_Rsvd:  %5lu\n"
 254                         "Hugepagesize:    %5lu kB\n",
 255                         nr_huge_pages,
 256                         free_huge_pages,
 257                         resv_huge_pages,
 258                         HPAGE_SIZE/1024);
 259 }
 260
 261 int hugetlb_report_node_meminfo(int nid, char *buf)
 262 {
 263         return sprintf(buf,
 264                 "Node %d HugePages_Total: %5u\n"
 265                 "Node %d HugePages_Free:  %5u\n",
 266                 nid, nr_huge_pages_node[nid],
 267                 nid, free_huge_pages_node[nid]);
 268 }
 269
 270 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 271 unsigned long hugetlb_total_pages(void)
 272 {
 273         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
 274 }
 275
 276 /*
 277  * We cannot handle pagefaults against hugetlb pages at all.  They cause
 278  * handle_mm_fault() to try to instantiate regular-sized pages in the
 279  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
 280  * this far.
 281  */
 282 static struct page *hugetlb_nopage(struct vm_area_struct *vma,
 283                                 unsigned long address, int *unused)
 284 {
 285         BUG();
 286         return NULL;
 287 }
 288
 289 struct vm_operations_struct hugetlb_vm_ops = {
 290         .nopage = hugetlb_nopage,
 291 };
 292
 293 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
 294                                 int writable)
 295 {
 296         pte_t entry;
 297
 298         if (writable) {
 299                 entry =
 300                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
 301         } else {
 302                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 303         }
 304         entry = pte_mkyoung(entry);
 305         entry = pte_mkhuge(entry);
 306
 307         return entry;
 308 }
 309
 310 static void set_huge_ptep_writable(struct vm_area_struct *vma,
 311                                    unsigned long address, pte_t *ptep)
 312 {
 313         pte_t entry;
 314
 315         entry = pte_mkwrite(pte_mkdirty(*ptep));
 316         ptep_set_access_flags(vma, address, ptep, entry, 1);
 317         update_mmu_cache(vma, address, entry);
 318         lazy_mmu_prot_update(entry);
 319 }
 320
 321
 322 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 323                             struct vm_area_struct *vma)
 324 {
 325         pte_t *src_pte, *dst_pte, entry;
 326         struct page *ptepage;
 327         unsigned long addr;
 328         int cow;
 329
 330         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 331
 332         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 333                 src_pte = huge_pte_offset(src, addr);
 334                 if (!src_pte)
 335                         continue;
 336                 dst_pte = huge_pte_alloc(dst, addr);
 337                 if (!dst_pte)
 338                         goto nomem;
 339                 spin_lock(&dst->page_table_lock);
 340                 spin_lock(&src->page_table_lock);
 341                 if (!pte_none(*src_pte)) {
 342                         if (cow)
 343                                 ptep_set_wrprotect(src, addr, src_pte);
 344                         entry = *src_pte;
 345                         ptepage = pte_page(entry);
 346                         get_page(ptepage);
 347                         add_mm_counter(dst, file_rss, HPAGE_SIZE / PAGE_SIZE);
 348                         set_huge_pte_at(dst, addr, dst_pte, entry);
 349                 }
 350                 spin_unlock(&src->page_table_lock);
 351                 spin_unlock(&dst->page_table_lock);
 352         }
 353         return 0;
 354
 355 nomem:
 356         return -ENOMEM;
 357 }
 358
 359 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 360                             unsigned long end)
 361 {
 362         struct mm_struct *mm = vma->vm_mm;
 363         unsigned long address;
 364         pte_t *ptep;
 365         pte_t pte;
 366         struct page *page;
 367         struct page *tmp;
 368         LIST_HEAD(page_list);
 369
 370         WARN_ON(!is_vm_hugetlb_page(vma));
 371         BUG_ON(start & ~HPAGE_MASK);
 372         BUG_ON(end & ~HPAGE_MASK);
 373
 374         spin_lock(&mm->page_table_lock);
 375
 376         /* Update high watermark before we lower rss */
 377         update_hiwater_rss(mm);
 378
 379         for (address = start; address < end; address += HPAGE_SIZE) {
 380                 ptep = huge_pte_offset(mm, address);
 381                 if (!ptep)
 382                         continue;
 383
 384                 pte = huge_ptep_get_and_clear(mm, address, ptep);
 385                 if (pte_none(pte))
 386                         continue;
 387
 388                 page = pte_page(pte);
 389                 list_add(&page->lru, &page_list);
 390                 add_mm_counter(mm, file_rss, (int) -(HPAGE_SIZE / PAGE_SIZE));
 391         }
 392
 393         spin_unlock(&mm->page_table_lock);
 394         flush_tlb_range(vma, start, end);
 395         list_for_each_entry_safe(page, tmp, &page_list, lru) {
 396                 list_del(&page->lru);
 397                 put_page(page);
 398         }
 399 }
 400
 401 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 402                           unsigned long end)
 403 {
 404         /*
 405          * It is undesirable to test vma->vm_file as it should be non-null
 406          * for valid hugetlb area. However, vm_file will be NULL in the error
 407          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
 408          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
 409          * to clean up. Since no pte has actually been setup, it is safe to
 410          * do nothing in this case.
 411          */
 412         if (vma->vm_file) {
 413                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 414                 __unmap_hugepage_range(vma, start, end);
 415                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 416         }
 417 }
 418
 419 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
 420                         unsigned long address, pte_t *ptep, pte_t pte)
 421 {
 422         struct page *old_page, *new_page;
 423         int avoidcopy;
 424
 425         old_page = pte_page(pte);
 426
 427         /* If no-one else is actually using this page, avoid the copy
 428          * and just make the page writable */
 429         avoidcopy = (page_count(old_page) == 1);
 430         if (avoidcopy) {
 431                 set_huge_ptep_writable(vma, address, ptep);
 432                 return VM_FAULT_MINOR;
 433         }
 434
 435         page_cache_get(old_page);
 436         new_page = alloc_huge_page(vma, address);
 437
 438         if (!new_page) {
 439                 page_cache_release(old_page);
 440                 return VM_FAULT_OOM;
 441         }
 442
 443         spin_unlock(&mm->page_table_lock);
 444         copy_huge_page(new_page, old_page, address);
 445         spin_lock(&mm->page_table_lock);
 446
 447         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
 448         if (likely(pte_same(*ptep, pte))) {
 449                 /* Break COW */
 450                 set_huge_pte_at(mm, address, ptep,
 451                                 make_huge_pte(vma, new_page, 1));
 452                 /* Make the old page be freed below */
 453                 new_page = old_page;
 454         }
 455         page_cache_release(new_page);
 456         page_cache_release(old_page);
 457         return VM_FAULT_MINOR;
 458 }
 459
 460 int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
 461                         unsigned long address, pte_t *ptep, int write_access)
 462 {
 463         int ret = VM_FAULT_SIGBUS;
 464         unsigned long idx;
 465         unsigned long size;
 466         struct page *page;
 467         struct address_space *mapping;
 468         pte_t new_pte;
 469
 470         mapping = vma->vm_file->f_mapping;
 471         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
 472                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 473
 474         /*
 475          * Use page lock to guard against racing truncation
 476          * before we get page_table_lock.
 477          */
 478 retry:
 479         page = find_lock_page(mapping, idx);
 480         if (!page) {
 481                 if (hugetlb_get_quota(mapping))
 482                         goto out;
 483                 page = alloc_huge_page(vma, address);
 484                 if (!page) {
 485                         hugetlb_put_quota(mapping);
 486                         ret = VM_FAULT_OOM;
 487                         goto out;
 488                 }
 489                 clear_huge_page(page, address);
 490
 491                 if (vma->vm_flags & VM_SHARED) {
 492                         int err;
 493
 494                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
 495                         if (err) {
 496                                 put_page(page);
 497                                 hugetlb_put_quota(mapping);
 498                                 if (err == -EEXIST)
 499                                         goto retry;
 500                                 goto out;
 501                         }
 502                 } else
 503                         lock_page(page);
 504         }
 505
 506         spin_lock(&mm->page_table_lock);
 507         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 508         if (idx >= size)
 509                 goto backout;
 510
 511         ret = VM_FAULT_MINOR;
 512         if (!pte_none(*ptep))
 513                 goto backout;
 514
 515         add_mm_counter(mm, file_rss, HPAGE_SIZE / PAGE_SIZE);
 516         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
 517                                 && (vma->vm_flags & VM_SHARED)));
 518         set_huge_pte_at(mm, address, ptep, new_pte);
 519
 520         if (write_access && !(vma->vm_flags & VM_SHARED)) {
 521                 /* Optimization, do the COW without a second fault */
 522                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
 523         }
 524
 525         spin_unlock(&mm->page_table_lock);
 526         unlock_page(page);
 527 out:
 528         return ret;
 529
 530 backout:
 531         spin_unlock(&mm->page_table_lock);
 532         hugetlb_put_quota(mapping);
 533         unlock_page(page);
 534         put_page(page);
 535         goto out;
 536 }
 537
 538 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 539                         unsigned long address, int write_access)
 540 {
 541         pte_t *ptep;
 542         pte_t entry;
 543         int ret;
 544         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
 545
 546         ptep = huge_pte_alloc(mm, address);
 547         if (!ptep)
 548                 return VM_FAULT_OOM;
 549
 550         /*
 551          * Serialize hugepage allocation and instantiation, so that we don't
 552          * get spurious allocation failures if two CPUs race to instantiate
 553          * the same page in the page cache.
 554          */
 555         mutex_lock(&hugetlb_instantiation_mutex);
 556         entry = *ptep;
 557         if (pte_none(entry)) {
 558                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
 559                 mutex_unlock(&hugetlb_instantiation_mutex);
 560                 return ret;
 561         }
 562
 563         ret = VM_FAULT_MINOR;
 564
 565         spin_lock(&mm->page_table_lock);
 566         /* Check for a racing update before calling hugetlb_cow */
 567         if (likely(pte_same(entry, *ptep)))
 568                 if (write_access && !pte_write(entry))
 569                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
 570         spin_unlock(&mm->page_table_lock);
 571         mutex_unlock(&hugetlb_instantiation_mutex);
 572
 573         return ret;
 574 }
 575
 576 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 577                         struct page **pages, struct vm_area_struct **vmas,
 578                         unsigned long *position, int *length, int i)
 579 {
 580         unsigned long pfn_offset;
 581         unsigned long vaddr = *position;
 582         int remainder = *length;
 583
 584         spin_lock(&mm->page_table_lock);
 585         while (vaddr < vma->vm_end && remainder) {
 586                 pte_t *pte;
 587                 struct page *page;
 588
 589                 /*
 590                  * Some archs (sparc64, sh*) have multiple pte_ts to
 591                  * each hugepage.  We have to make * sure we get the
 592                  * first, for the page indexing below to work.
 593                  */
 594                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
 595
 596                 if (!pte || pte_none(*pte)) {
 597                         int ret;
 598
 599                         spin_unlock(&mm->page_table_lock);
 600                         ret = hugetlb_fault(mm, vma, vaddr, 0);
 601                         spin_lock(&mm->page_table_lock);
 602                         if (ret == VM_FAULT_MINOR)
 603                                 continue;
 604
 605                         remainder = 0;
 606                         if (!i)
 607                                 i = -EFAULT;
 608                         break;
 609                 }
 610
 611                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
 612                 page = pte_page(*pte);
 613 same_page:
 614                 if (pages) {
 615                         get_page(page);
 616                         pages[i] = page + pfn_offset;
 617                 }
 618
 619                 if (vmas)
 620                         vmas[i] = vma;
 621
 622                 vaddr += PAGE_SIZE;
 623                 ++pfn_offset;
 624                 --remainder;
 625                 ++i;
 626                 if (vaddr < vma->vm_end && remainder &&
 627                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
 628                         /*
 629                          * We use pfn_offset to avoid touching the pageframes
 630                          * of this compound page.
 631                          */
 632                         goto same_page;
 633                 }
 634         }
 635         spin_unlock(&mm->page_table_lock);
 636         *length = remainder;
 637         *position = vaddr;
 638
 639         return i;
 640 }
 641
 642 void hugetlb_change_protection(struct vm_area_struct *vma,
 643                 unsigned long address, unsigned long end, pgprot_t newprot)
 644 {
 645         struct mm_struct *mm = vma->vm_mm;
 646         unsigned long start = address;
 647         pte_t *ptep;
 648         pte_t pte;
 649
 650         BUG_ON(address >= end);
 651         flush_cache_range(vma, address, end);
 652
 653         spin_lock(&mm->page_table_lock);
 654         for (; address < end; address += HPAGE_SIZE) {
 655                 ptep = huge_pte_offset(mm, address);
 656                 if (!ptep)
 657                         continue;
 658                 if (!pte_none(*ptep)) {
 659                         pte = huge_ptep_get_and_clear(mm, address, ptep);
 660                         pte = pte_mkhuge(pte_modify(pte, newprot));
 661                         set_huge_pte_at(mm, address, ptep, pte);
 662                         lazy_mmu_prot_update(pte);
 663                 }
 664         }
 665         spin_unlock(&mm->page_table_lock);
 666
 667         flush_tlb_range(vma, start, end);
 668 }
 669
 670 struct file_region {
 671         struct list_head link;
 672         long from;
 673         long to;
 674 };
 675
 676 static long region_add(struct list_head *head, long f, long t)
 677 {
 678         struct file_region *rg, *nrg, *trg;
 679
 680         /* Locate the region we are either in or before. */
 681         list_for_each_entry(rg, head, link)
 682                 if (f <= rg->to)
 683                         break;
 684
 685         /* Round our left edge to the current segment if it encloses us. */
 686         if (f > rg->from)
 687                 f = rg->from;
 688
 689         /* Check for and consume any regions we now overlap with. */
 690         nrg = rg;
 691         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 692                 if (&rg->link == head)
 693                         break;
 694                 if (rg->from > t)
 695                         break;
 696
 697                 /* If this area reaches higher then extend our area to
 698                  * include it completely.  If this is not the first area
 699                  * which we intend to reuse, free it. */
 700                 if (rg->to > t)
 701                         t = rg->to;
 702                 if (rg != nrg) {
 703                         list_del(&rg->link);
 704                         kfree(rg);
 705                 }
 706         }
 707         nrg->from = f;
 708         nrg->to = t;
 709         return 0;
 710 }
 711
 712 static long region_chg(struct list_head *head, long f, long t)
 713 {
 714         struct file_region *rg, *nrg;
 715         long chg = 0;
 716
 717         /* Locate the region we are before or in. */
 718         list_for_each_entry(rg, head, link)
 719                 if (f <= rg->to)
 720                         break;
 721
 722         /* If we are below the current region then a new region is required.
 723          * Subtle, allocate a new region at the position but make it zero
 724          * size such that we can guarentee to record the reservation. */
 725         if (&rg->link == head || t < rg->from) {
 726                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
 727                 if (nrg == 0)
 728                         return -ENOMEM;
 729                 nrg->from = f;
 730                 nrg->to   = f;
 731                 INIT_LIST_HEAD(&nrg->link);
 732                 list_add(&nrg->link, rg->link.prev);
 733
 734                 return t - f;
 735         }
 736
 737         /* Round our left edge to the current segment if it encloses us. */
 738         if (f > rg->from)
 739                 f = rg->from;
 740         chg = t - f;
 741
 742         /* Check for and consume any regions we now overlap with. */
 743         list_for_each_entry(rg, rg->link.prev, link) {
 744                 if (&rg->link == head)
 745                         break;
 746                 if (rg->from > t)
 747                         return chg;
 748
 749                 /* We overlap with this area, if it extends futher than
 750                  * us then we must extend ourselves.  Account for its
 751                  * existing reservation. */
 752                 if (rg->to > t) {
 753                         chg += rg->to - t;
 754                         t = rg->to;
 755                 }
 756                 chg -= rg->to - rg->from;
 757         }
 758         return chg;
 759 }
 760
 761 static long region_truncate(struct list_head *head, long end)
 762 {
 763         struct file_region *rg, *trg;
 764         long chg = 0;
 765
 766         /* Locate the region we are either in or before. */
 767         list_for_each_entry(rg, head, link)
 768                 if (end <= rg->to)
 769                         break;
 770         if (&rg->link == head)
 771                 return 0;
 772
 773         /* If we are in the middle of a region then adjust it. */
 774         if (end > rg->from) {
 775                 chg = rg->to - end;
 776                 rg->to = end;
 777                 rg = list_entry(rg->link.next, typeof(*rg), link);
 778         }
 779
 780         /* Drop any remaining regions. */
 781         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 782                 if (&rg->link == head)
 783                         break;
 784                 chg += rg->to - rg->from;
 785                 list_del(&rg->link);
 786                 kfree(rg);
 787         }
 788         return chg;
 789 }
 790
 791 static int hugetlb_acct_memory(long delta)
 792 {
 793         int ret = -ENOMEM;
 794
 795         spin_lock(&hugetlb_lock);
 796         if ((delta + resv_huge_pages) <= free_huge_pages) {
 797                 resv_huge_pages += delta;
 798                 ret = 0;
 799         }
 800         spin_unlock(&hugetlb_lock);
 801         return ret;
 802 }
 803
 804 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
 805 {
 806         long ret, chg;
 807
 808         chg = region_chg(&inode->i_mapping->private_list, from, to);
 809         if (chg < 0)
 810                 return chg;
 811         ret = hugetlb_acct_memory(chg);
 812         if (ret < 0)
 813                 return ret;
 814         region_add(&inode->i_mapping->private_list, from, to);
 815         return 0;
 816 }
 817
 818 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
 819 {
 820         long chg = region_truncate(&inode->i_mapping->private_list, offset);
 821         hugetlb_acct_memory(freed - chg);
 822 }